Monocular 3D Vehicle Detection on Roadside based on Foreground Depth Fusion

With the rapid development of deep learning techniques, monocular 3D vehicle detection has become a hot research field. Compared to traditional 2D vehicle detection, 3D detection provides more detailed information about vehicles, such as their dimensions (length, width, height), orientation angles, and more. However, this also places higher demands on the algorithms used.We present a two-stage monocular 3D vehicle detection framework that integrates foreground depth fusion. In the initial stage, a 2D detector is employed to generate 2D bounding boxes from the input images, followed by hierarchical annotation of different detection targets. In the subsequent stage, we fuse the foreground depth and hierarchical information to derive accurate 3D information for the identified objects. Experimental evaluation on the DAIR-V2X dataset showcases superior performance of our network compared to other approaches in handling occlusion issues.